Evolutionary algorithms for predicting aboveground carbon stocks inmopane woodlands in Mozambique
Genetic Algorithm andRandom Forest (GARF);Genetic Programming (GP);symbolic regression; drytropical forests; machinelearning
Tropical forests are crucial for global climate regulation and carbon cycling. Mopane wood-lands, a tropical dry forest covering southern Africa, feature high ecological-socioeconomicimportance. In Mozambique, charcoal production is a major driver of Mopane degradationand aboveground carbon (AGC) loss. Accurate AGC estimation is essential for climate mitiga-tion strategies. We applied machine learning techniques to predict stand-level AGC inMopane woodlands across Mabalane and Chicualacuala districts, Gaza Province. Two evolu-tionary algorithms were tested: (1) a hybrid Genetic Algorithm and Random Forest (GARF),and (2) Genetic Programming (GP) using symbolic regression. In total, 139 predictor variableswere derived from remote sensing, biophysical, and bioclimatic datasets. Field data included114 cluster plots. Both algorithms reduced the dataset by 95.6%. Observed AGC rangedfrom 1.313 to 28.476 MgC ha−1. GARF predictions ranged from 2.910 to 19.459 MgC ha−1(nRMSE ¼ 0.427; MBE ¼ 0.08), while GP showed a wider predictive range (1.721–23.503MgC ha−1; nRMSE ¼ 0.428; MBE ¼ 2.731 10−17). GARF relied on optical and bioclimatic vari-ables, whereas GP operated independently of variable type. Both approaches were effectivefor feature selection and AGC prediction. However, GP produced a more interpretablemodel, offering advantages for replication and use in operational carbon inventories.